相较于英伟达、AMD：国产显卡为何不好用 中国工程院院士释疑！

Kuai Technology reported on November 2 that according to Zheng Weimin, an academician of the Chinese Academy of Engineering and a professor of the Department of Computer Science and Technology of Tsinghua University, the core challenge facing the current domestic computing industry has never been just the performance of the chip hardware itself, but also stuck in the maturity of the software ecosystem and the actual comprehensive use cost on the user side, which have been underestimated for a long time. In response to the common pain points that many developers and enterprise users have complained about, the actual use experience of domestic graphics cards is not as good as that of AMD and NVIDIA products. Zheng Weimin analyzed that with the rapid popularization of new forms of applications such as intelligent agents, the demand for large model Token calls has recently shown explosive growth, and the rigid demand for computing power in the entire market continues to rise rapidly. At present, the hardware paper parameters and terminal purchase prices of some domestic computing power chips are actually quite competitive in the global market. However, when they are actually implemented in actual production environments, they will generally encounter practical problems such as complicated software adaptation processes and high secondary development costs, which directly reduces the comprehensive use experience of the products. Zheng Weimin specifically pointed out that domestic computing power chips are not incapable of achieving high performance. As long as sufficient resources are invested in targeted software adaptation and optimization, they can achieve very good operating performance. However, currently, domestic chips of different manufacturers and models often need to be adapted and optimized separately, and the entire process consumes a large amount of repeated development resources. Compared with the purchase price of the chip itself, the subsequent high usage cost is now the most critical factor affecting the selection of enterprise users. He also announced relevant measured data. The research team ran typical training and inference task tests on four types of mainstream domestic computing platforms. After completing targeted software optimization, the corresponding performance can reach 86% to 98% of the corresponding performance of NVIDIA A100. The actual hardware performance gap has been narrowed to a very small range. Zheng Weimin concluded that research and development on the chip hardware side is of course very important, but the construction of supporting software ecosystem is also a core link that cannot be ignored in industrial development. Compared with the mature computing power industry chain in the United States, the current hardware performance of domestic chips has laid a pretty good foundation, but the maturity of the overall software ecosystem is still not small.